Modified LIBERO Simulation Benchmark With High-Resolution Images

Presenting a modified version of the LIBERO simulation benchmark dataset, created for the Cosmos Policy project. It was used to train the Cosmos-Policy-LIBERO-Predict2-2B checkpoint. The dataset includes higher-resolution images at 256x256 pixels and has filtered transitions with 'no-op' actions.

Use Cases

• Train reinforcement learning agents for robotics tasks using the modified LIBERO simulation benchmark.
• Fine-tune computer vision models on the provided 256x256 pixel images for scene understanding in simulated environments.
• Benchmark policy learning algorithms on a dataset with filtered 'no-op' action transitions.

Strengths

• Images are provided at a higher resolution of 256x256 pixels compared to the original 128x128.
• The dataset is a modified version of the established LIBERO simulation benchmark.
• Transitions containing 'no-op' (zero) actions have been filtered.

Limitations

• The dataset's total size, number of rows, and specific column structure are unknown.
• The license terms for use and redistribution are unspecified.
• The dataset is derived from a simulation, which may not fully capture real-world physics and noise.

Provenance

Source

nvidia

Collection Method

Modified from the original LIBERO simulation benchmark dataset.

Freshness

The dataset was last updated on 2026-02-05.